Professional Documents
Culture Documents
2 Linear Functionals 23
2.1 The Riesz representation theorem . . . . . . . . . . . . . . . . . . . . . . . . 23
2.2 The Riesz Representation Theorem for Hilbert spaces . . . . . . . . . . . . . 27
2.3 The Hahn-Banach Theorems . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.4 A few results about convex sets . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.5 The dual of a topological vector space . . . . . . . . . . . . . . . . . . . . . . 33
2.5.1 The weak -topology . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.5.2 The dual of a normed vector space . . . . . . . . . . . . . . . . . . . 35
3
4 CONTENTS
5 C algebras 67
5.1 The definition of C -algebras . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
5.2 Commutative C -algebras . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
5.3 C -algebras as algebras of operators . . . . . . . . . . . . . . . . . . . . . . . 70
5.4 Functional calculus for normal operators . . . . . . . . . . . . . . . . . . . . 71
A Background results 83
A.1 Zorns lemma . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
Chapter 1
The study of differential and integral equations (understanding the solution set).
Definition. A vector space over C (or R) is a set V endowed with an addition and a scalar
multiplication with the following properties
5
6 CHAPTER 1. TOPOLOGICAL VECTOR SPACES
Definition. If V and W are vector spaces, a map T : V W is called a linear map (or
linear operator) if for every scalars and and every vectors x, y V ,
T (x + y) = T x + T y.
and X are in T
Definition. A map f : X Y is called continuous if for every open set U Y , the set
f 1 (U ) is open in X.
A map is called a homeomorphism if it is invertible and both the map and its inverse are
continuous. A topological space X is called Hausdorff if for every x, y X, there are open
sets U1 and U2 such that x U1 , y U2 and U1 U2 = .
A neighborhood of x is an open set containing x. A system of neighborhoods of x is a
family of neighborhoods of x such that for every neighborhood of x there is a member of this
family inside it.
A subset C X is closed if X\C is open. A subset K X is called compact if every
covering of K by open sets has a finite subcover. The closure of a set is the smallest closed
set that contains it. If A is a set then A denotes its closure. The interior of a set is the
largest open set contained in it. We denote by int(A) the interior of A.
Definition. A topological vector space over the field K (which is either C or R) is a vector
space X endowed with a topology such that every point is closed and with the property that
both addition and scalar multiplication,
+ : X X X and : K X X,
are continuous.
1.3. BASIC PROPERTIES OF TOPOLOGICAL VECTOR SPACES 7
Example. Rn is an example of a finite dimensional topological vector space, while C([0, 1])
is an example of an infinite dimensional vector space.
A subset E of a topological vector space is called bounded if for every neighborhood U
of 0 there is a number s > 0 such that E tU for every t > s.
A topological vector space is called locally convex if every point has a system of neigh-
borhoods that are convex.
(K + U ) (C + U ) = .
Proof. Applying twice Proposition 1.3.2 twice we we deduce that for every neighborhood W
of 0 there is an open symmetric neighborhood U of 0 such that U + U + U + U W . Since
the topology is translation invariant, it means that for every neighborhood W of a point x
there is an open symmetric neighborhood of 0, Ux , such that x + Ux + Ux + Ux + Ux W .
Now let x K and W = X\C. Then x + Ux + Ux + Ux + Ux X\C, and since Ux is
symmetric, (x + Ux + Ux ) (C + Ux + Ux ) = .
Since K is compact, there are finitely many points x1 , x2 , . . . , xk such that K (x1 +
Ux1 ) (x2 + Ux2 ) (xk + Uxk ). Set U = Ux1 Ux2 Uxk . Then
X =
n=1 rn U.
h , i:V V C
hy, xi = hx, yi
hx, yi = xT y.
hz, wi = zT w.
Example. The space C([0, 1]) of continuous functions f : [0, 1] C with the inner product
Z 1
hf, gi = f (t)g(t)dt.
0
and the distance between two elements is defined to be kx yk. Two elements, x and y, are
called orthogonal if
< x, y >= 0.
The norm completely determines the inner product by the polarization identity which in
the case of vector spaces over R is
1
hx, yi = (kx + yk2 kx yk2 )
4
and in the case of vector spaces over C is
1
hx, yi = (kx + yk2 kx yk2 + ikx + iyk2 ikx iyk2 ).
4
Note that we also have the parallelogram identity
Proposition 1.4.1. The norm induced by the inner product has the following properties:
a) kxk = ||kxk,
b) (the Cauchy-Schwarz inequality) | hx, yi | kxkkyk,
c) (the Minkowski inequality aka the triangle inequality) kx + yk kxk + kyk.
10 CHAPTER 1. TOPOLOGICAL VECTOR SPACES
Proof. Part a) follows easily from the definition. For b), choose of absolute value 1 such
that hx, yi > 0. Let also t be a real parameter. We have
and the fact that this is less than or equal to zero is equivalent to the Cauchy-Schwarz
inequality.
For c) we use the Cauchy-Schwarz inequality and compute
Proposition 1.4.2. The vector space V endowed with the inner product has a natural
topological vector space structure in which the open sets are arbitrary unions of balls of the
form
Proof. The continuity of addition follows from the triangle inequality. The continuity of the
scalar multiplication is straightforward.
Definition. A Hilbert space is a vector space H endowed with an inner product, which is
complete, in the sense that if xn is a sequence of points in H that satisfies the condition
kxn xm k 0 for m, n , then there is an element x H such that kxn xk 0.
Using the Cauchy-Schwarz inequality, we see that strong convergence implies weak con-
vergence.
Definition. The dimension of a Hilbert space is the smallest cardinal number of a set of
elements whose finite linear combinations are everywhere dense in the space.
We will only be concerned with Hilbert spaces of either finite or countable dimension.
1.4. HILBERT SPACES 11
Definition. An orthonormal basis for a Hilbert space is a set of unit vectors that are pairwise
orthogonal and such that the linear combinations of these elements are dense in the Hilbert
space.
Proposition 1.4.3. Every separable Hilbert space has an orthonormal basis.
Proof. Consider a countable dense set in the Hilbert space and apply the Gram-Schmidt
process to it.
From now on we will only be concerned with separable Hilbert spaces.
Theorem 1.4.1. If en , n 1 is an orthonormal basis of the Hilbert space H, then:
a) Every element x H can be written uniquely as
X
x= cn en ,
n
Note that
N
X N
X
2
k hx, en i en k = | hx, en i |2 ,
n=1 n=1
P
and so Bessels inequality shows that n1 hx, en i en converges.
Given that the set of vectors of the form N
P
n=1 cn en is dense in the Hilbert space, and
that such a sum best approximates x if cn =< x, en >, we conclude that
X
x= hx, en i en .
n=1
Example. The Hardy space on the unit disk H 2 (D). It consists of the holomorphic functions
on the unit disk for which
Z 2 1/2
1 i 2
sup |f (re )| d
0<r<1 2 0
is finite. This quantity is the norm of H 2 (D); it comes from an inner product. An orthonormal
basis consists of the monomials 1, z, z 2 , z 3 , . . ..
HL2 (C, ~ )
x = (x1 , x2 , x3 , . . .)
We set
X
< x, y >= x n yn .
n=1
Theorem 1.4.2. Every two Hilbert spaces (over the same field of scalars) of the same
dimension are isometrically isomorphic.
Proof. Let (en )n and (e0n )n be orthonormal bases of the first, respectively second space. The
map
X X
cn en 7 cn e0n
n n
preserves the norm. The uniqueness of writing an element in an orthonormal basis implies
that this map is linear.
Corollary 1.4.1. Every separable Hilbert space over C is isometrically isomorphic to either
Cn for some n or to l2 (C). Every separable Hilbert space over R is isometrically isomorphic
to either Rn for some n or to l2 (R).
and because convergence in norm implies the convergence of coefficients, it follows that the
limit of a sequence of elements in E is also a linear combination of e1 , e2 , . . . , eN , hence is in
E.
However, if the Hilbert space H is infinite dimensional, then there are subspaces which are
not closed. For example if e1 , e2 , e3 , . . . is an orthonormal basis, then the linear combinations
of these basis elements define a subspace which is dense, but not closed because it is not the
whole space.
Definition. We say that an element x is orthogonal to a subspace E if xe for every e E.
The orthogonal complement of a subspace E is
hx + y, ei = hx, ei + hy, ei = 0,
which shows that E is a subspace. The fact that it is closed follows from the fact that strong
convergence implies weak convergence.
1.4. HILBERT SPACES 15
Theorem 1.4.3. (The decomposition theorem) If E is a closed subspace of the Hilbert space
H, then every x H can be written uniquely as x = y + z, where y E and z E .
Proof. (Following Riesz-Nagy) Consider y E as variable and consider the distances kxyk.
Let d be their infimum, and yn a sequence such that
kx yn k d.
Using it we obtain
yn + ym 2
kyn ym k2 = 2(kx yn k2 + kx ym k2 ) 4kx k
2
2(kx yn k2 + kx ym k2 ) 4d2 .
The last expression converges to 0 when m, n . This implies that yn is Cauchy, hence
convergent. Let y E be its limit. Then kx yk = d.
Set z = x y. We will show that z is orthogonal to E. For this, let y0 be an arbitrary
element of E. Then for every C,
| hx y, y0 i |2
0.
ky0 k2
Exercise. Show that every nonempty closed convex subset of H contains a unique element
of minimal norm.
16 CHAPTER 1. TOPOLOGICAL VECTOR SPACES
k k : X [0, )
Definition. A Banach space is a normed vector space that is complete, namely in which
every Cauchy sequence of elements converges.
Example. Every Hilbert space is a Banach space. In fact, the necessary structure for a
Banach space to have an underlying Hilbert space structure (prove it!) is that the norm
satisfies
is a Banach space.
Example. Let p 1 be a real number. The space Cn endowed with the norm
is a Banach space.
Example. The space C([0, 1]) of continuous functions on [0, 1] is a Banach space with the
norm
kf k = sup |f (t)|.
t[0,1]
1.5. BANACH SPACES 17
is a Banach space. It is also separable. In general the Lp space over any measurable space
is a Banach space.
The space L (R) of functions that are bounded almost everywhere is also Banach. Here
two functions are identified if they coincide almost everywhere. The norm is defined by
The Sobolev spaces with 1 p < are separable. However, for p = , one defines the
norm to be
max kD f k ,
||k
1.6.1 Seminorms
Definition. A seminorm on a vector space is a function
k k : X [0, )
so p(x + y) = 0.
b) This is just a rewriting of the triangle inequality.
c) The fact that is balanced follows from kxk = |kkxk. For convexity, note that
Proof. a) Consider > 0 and let t = A (x) + , s = A (y) + . Then x/t and y/s are in A
and so is their convex combination
x+y t x s y
= + .
s+t s+t t s+t s
It follows that A (x + y) s + t = A (x) + A (y) + 2. Now pass to the limit 0.
b) follows from the definition and c) follows from a) and b).
For d) note that the inclusions B A C show that C A B . For the converse
inequalities, let x X and choose t, s such that C (x) < s < t. Then x/s C so A (x/s) 1
and A (x/t) < 1. Hence x/t B, so B (x) t. Vary t to obtain B C .
A family of seminorms P on a vector space is called separating if for every x 6= y, there
is a seminorm p P such that p(x y) > 0.
Proposition 1.6.3. Suppose X has a system of neighborhoods of 0 that are convex and
balanced. Associate so each open set V in this system of neighborhoods its Minkowski
functional V . Then V = {x X | V (x) < 1}, and the family of functionals V defined for
all such V s is a separating family of continuous functionals.
Proof. If x V then x/t is still in V for some t > 1, so V (x) < 1. If x 6 V , then x/t V
implies t 1 because V is balanced and convex. This proves that V = {x X | V (x) < 1}.
By Proposition 1.6.2, V is a seminorm for all V . Applying Proposition 1.6.1 b) we have
that for every > 0 if x y V then
|V (x) V (y)| V (x y) < ,
which proves the continuity of V at x. Finally, V is separating because X is Hausdorff.
Theorem 1.6.1. Suppose P is a separating family of seminorms on a vector space X.
Associate to each p P and to each positive integer n the set
B1/n,p = {x | p(x) < 1/n}.
Let V be the set of all finite intersections of such sets. Then V is a system of convex,
balanced, absorbing neighborhoods of 0, which defines a topology on X and turns X into a
topological vector space such that every p P is continuous and a set A is bounded if and
only if p|A is bounded for all p.
Proof. Proposition 1.6.1 implies that each set B1/n,p is convex and balanced, and hence so are
the sets in V. Consider all translates of sets in V, and let the open sets be arbitrary unions
of such translates. We thus obtain a topology on X. Because the family is separating,
the topology is Hausdorff. We need to check that addition and scalar multiplication are
continuous.
Let U be a neighborhood of 0 and let
B1/n1 ,p1 B1/n2 ,p2 B1/nk ,pk U.
Set
V = B1/2n1 ,p1 B1/2n2 ,p2 B1/2nk ,pk .
20 CHAPTER 1. TOPOLOGICAL VECTOR SPACES
Since every open neighborhood of zero contains such an open subset, A is bounded.
Definition. A Frechet space is a topological vector space with the properties that
it is Hausdorff
kf kk = sup |f (x)|.
kxkleqk
Example. Let D be an open subset of the complex plane. There is a sequence of compact
sets K1 K2 K3 D whose union is D. Let H(D) be the space of holomorphic
functions on D endowed with the seminorms
kf kk = sup{f (z) | z Kj }.
Theorem 1.6.2. A topological vector space X has a norm that induces the topology if and
only if there is a convex bounded neighborhood of the origin.
Proof. If a norm exists, then the open unit ball centered at the origin is convex and bounded.
For the converse, assume V is such a neighborhood. By Theorem 1.3.1, V contains a
convex balanced neighborhood, which is also bounded. Let k k be the Minkowski functional
of this neighborhood. By Proposition 1.3.5, the rU , r 0, is a family of neighborhoods of 0.
Moreover, because U is bounded, for every x there is r > 0 such that x 6 rU . Then kxk r
so kxk = 0 if and only if x = 0. Thus k k is a norm and the topology is induced by this
norm.
22 CHAPTER 1. TOPOLOGICAL VECTOR SPACES
Chapter 2
Linear Functionals
for all n 0?
We ask the more general problem, if there is a measure on R such that tn L1 () for
all n and
Z
sn = tn d(t).
It is easy to see that not all such sequences are moments. Set
Xn
p(t) = aj tj .
j=0
Then
Z n Z
X
2
0 |p(t)| d(t) = tj+k d(t)aj ak
j,k=0
n
X
= sj+k aj ak .
j,k=0
23
24 CHAPTER 2. LINEAR FUNCTIONALS
Theorem 2.1.1. (M. Riesz) Let X be a linear space over R and C X a cone, meaning
that if x, y C and t > 0 then x + y C and tx C. Assume moreover that the cone is
proper, meaning that C (C) = {0} and define the order x y if and only if y x C.
Let Y X be a subspace and let 0 : Y R be a linear functional such that 0 (y) 0
for all y Y C. Supose that for every x X there is u Y C and t > 0 such that
tu x C. Then there is a linear functional : X X such that |Y = 0 and (x) 0
for all x C.
Proof. First, let us assume X = Y + Rx with x 6 Y . Suppose we are able to set (x) = .
If y Y such that y + tx C then we should have (y) + t(x) 0 so 0 (y) + t 0. All
we need to show is that inf{0 (y)|x + y C} is not . Write x = u c with u Y C
and c C. Then y u C, so 0 (y) 0 (u). We set
For the general case, apply Zorns Lemma. Consider the set M of functionals : Z R
such that Y Z X, positive, and |Y = 0 . Order it by
Theorem 2.1.2. (F. Riesz) Let : C([0, 1]) R be a positive linear functional. Then
there is a unique positive measure on [0, 1] such that
Z 1
(f ) = f (t)d(t). (2.1.2)
0
Proof. We use the theorem of M. Riesz. Let B([0, 1]) be the space of bounded functions on
[0, 1]. We have Set X = B([0, 1]) and Y = C([0, 1]). The conditions of Theorem 2.1.1 are
satisfied, because every bounded function is the difference between a continuous bounded
function and a positive function. Hence there is a positive linear functional : X R such
that |C([0, 1]) = . Define the monotone increasing function F : [0, 1] R such that
F (t) = ([0,t] ).
2.1. THE RIESZ REPRESENTATION THEOREM 25
(A) = (A ),
Proof. Denote by R[x] the real valued polynomial functions on R and by Cc (R) the contin-
uous functions with compact support. Consider
and
The fact that is a finite measure is proved as follows. Given an interval [m, m], let f
be compactly supported, such that
[m,m] f 1.
Then
([m, m]) = ([m,m] ) (f ) 0 (1) = s0 .
Let us now show that Z
(p) = p(t)d(t).
If p is an even degree polynomial with positive dominant coefficient, then it can be approxi-
mated from below by compactly supported continuous functions, and so using the positivity
of we conclude that for every such function f
Z
(p) (f ) = f (t)d(t).
By passing to the limit we find that
Z
(p) p(t)d(t).
Let q be a polynomial of even degree with dominant coefficient positive, whose degree is less
than p. Then for every a > 0,
Z
(p aq) (p aq)(t)d(t).
Said differently
Z Z
(p) p(t)d(t) a (q) q(t)d(t) .
This can only happen if
Z
(q) = q(t)d(t).
Varying p and q we conclude that this is true for every q with even degree and with positive
dominant coefficient. Since every polynomial can be written as the difference between two
such polynomials, the property is true for all polynomials.
2.2. THE RIESZ REPRESENTATION THEOREM FOR HILBERT SPACES 27
Again, (z, w) 7 Kz (w) is called the reproducing kernel of the Segal-Bargmann space.
Remark 2.2.1. Using the Cauchy-Schwarz inequality, we see that
|(x)| kzkkxk.
In fact it can be seen that the continuity of is equivalent to the existence of an inequality
of the form |(x)| Ckxk that holds for all x, where C is a fixed positive constant.
28 CHAPTER 2. LINEAR FUNCTIONALS
Remark 2.3.1. The functional p can be a seminorm, or more generally, a Minkowski func-
tional.
Y1 = {y + tx1 | y Y, t R}.
Because
we have
Define 1 : Y1 R, by
1 (y + tx1 ) = 0 (y) + t.
Proof. This is an easy consequence of the previous result. If we work with real vector spaces,
then because is linear, by changing x to x if necessary, we get |(x)| p(x) for all x.
If X is a complex vector space, note that a linear functional can be decomposed as
= Re + iIm. Then Re(ix) = Im(x), so (x) = Re(x) + iRe(ix). So the real part
determines the functional.
Apply the theorem to Re0 to obtain Re, and from it recover . Note that for every
x X, there is C, || = 1 such that (x) = |(x)|. Then
B = {x x0 a + a0 | x A, x0 A0 }.
It is not hard to check that B is convex, it is also absorbing. Consider the Minkowski
functional B . Because A and A0 are disjoint, a0 a 6 B. Hence B (a0 a) 1.
Set Y = R(a0 a) and define 0 ((a0 a)) = . Then 0 (y) B (y) for all y Y . By
the Hahn-Banach theorem, there is : X R such that (x) B (x) for all x X, and
also (a0 a) = 0 (a0 a) = 1.
If x B, then B (x) 1, so (x) 1. Hence if x A, x0 A0 , then
(x x0 a + a0 ) 1.
In other words
Theorem 2.3.4. (Hahn-Banach) Suppose A and B are disjoint, nonempty, convex sets in
a locally convex topological vector space X.
a) If A is open then there is a continuous linear functional on X and R such that
Re(x) < Re(y) for all x A, y Y.
b) If A is compact and B is closed then there is a continuous linear functional and 1 , 2 R
such that
Re(x) 1 < 2 Re(y)
for all x A and y B.
Proof. We just consider the real case. Because A is open, every point of A is internal. So
there is and such that A {x | (x) } and B {x | (x) }. Let x0 be a point in
A. Then (x x0 ) = (x) (x0 ) (x0 ). So there is an open neighborhood V of 0
such that (y) if y V , where = (x0 ).. Choosing V to be balanced, we conclude
that |(y)| for all y V . But this is the condition that is continuous.
We claim that because A is open (x) < for all x A. If not, let x be such that
(x) = . Given y A, there is a small t (0, 1) such that (x ty)/(1 t) A (because of
the continuity of addition and multiplication). Call this point z. Then x = ty + (1 t)z, and
because (y) and (z) are both less than or equal to , (y) = (z) = (x) = . Hence
is constant in a neighborhood of x. Consequently is constant in a neighborhood of 0, and
because the neighborhoods of 0 are absorbing, it is constant everywhere. This is impossible.
Hence a) is true.
For b) we use Proposition 1.3.3 to conclude that there is an open set U such that A +
U B + U = . By shrinking, we can make U balanced and convex. Let x0 U \{0} such
that (x0 ) = 6= 0. Then
sup (x) + sup (x), inf (y) inf (y).
xA xA+U yB yB+U
K0 = K 0 G K 0 6= .
It is not hard to see that K0 is also extremal. Hence we are in the conditions of Zorns
Lemma. We deduce that F has minimal elements.
Let Km be a minimal element; we claim it is a singleton. Arguing by contradiction,
let us assume that Km has two distinct points. By the Hahn-Banach Theorem there is a
continuous linear functional : X R such that (x) 6= (y). Let
= max (x).
xKm
Define
K1 = {y K0 | (y) = }.
Ext(K) 6= .
Theorem 2.4.2. (Milman) Let X be a locally convex topological vector space. Let K be a
compact set such that co(K) is also compact. Then every extreme point of co(K) lies in K.
Proof. Assume that some extreme point x0 co(K) is not in K. Then there is a convex
balanced neighborhood V of 0 in X such that
(x0 + V ) K = .
K j xj + V = xj + V .
co(K) co(K1 Kn ).
The opposite inclusion also holds, because Kj co(K) for every j. Hence
co(K) = co(K1 Kn ).
In particular,
x0 = t1 y1 + t2 y2 + + tn yn ,
P
where yj Kj and tj 0, tj = 1. But x0 is extremal in co(K), so x0 coincides with one
of the yj . Thus for some j,
x0 K j xj + V K + V ,
We will show an application of these results. Given a convex subset K of a vector space
X, and a vector space Y , a map T : K Y is called affine if for every x, y K and t [0, 1],
The result is about groups of affine transformations from X into itself. If X has a topological
vector space structure, a group G of affine transformations is called equicontinuous if for every
neighborhood V of 0 in X, there is a neighborhood U of 0 such that T (x) T (y) V for
every x, y K such that x y U and for every T G.
Theorem 2.4.3. (Kakutanis Fixed Point Theorem) Suppose that K is a nonempty compact
convex subset of a locally topological space X and G is an equicontinuous group of affine
transformations taking K into itself. Then there is x0 K such that T (x0 ) = x0 for all
T G.
2.5. THE DUAL OF A TOPOLOGICAL VECTOR SPACE 33
Proof. Let
F = {K 0 K | K 0 : nonempty, compact, convex , T (K 0 ) K 0 for all T G}.
Note that K F, so this family is nonempty. Order F by inclusion and note that if G is a
subfamily that is totally ordered, then because K is compact,
K0 = K 0 G K 0 6= .
Clearly T (K0 ) K0 , so the conditions of Zorns lemma are satisfied. It follows that F has
minimal elements. Let K0 be such a minimal element. We claim that it is a singleton.
Assume, to the contrary, that K0 contains x, y with x 6= y. Let V be a neighborhood of
0 such that x y 6 V , and let U be the neighborhood of 0 associated to U by the definition
of equicontinuity. Then for every T G, T (x) T (y) 6 U , for else, because T 1 G,
x y = T 1 (T (x)) T 1 (T (y)) V.
Set z = 21 (x + y). Then z K0 . Let
G(z) = {T (z) | T G}.
the G(z) is G-invariant, hence so is its closure K1 = G(z). Consequently, co(K1 ) is a
G-invariant, compact convex subset of K0 . The minimality of K0 implies
K0 = co(K1 ).
By the Krein-Milman Theorem (Theorem 2.4.1), K0 has extremal points. Applying Milmans
Theorem (Theorem 2.4.2), we deduce that every extremal point of K0 lies in K1 . Let p be
such a point.
Consider the set
S = {(T z, T x, T y) | T G} K0 K0 K0 .
Since p K1 = G(z), and K0 K0 is compact, there is a point (x1 , y1 ) K0 K0
such that (x0 , x1 , y1 ) S. Indeed, if this were not true, then every (x1 , y1 ) K0 K0
had a neighborhood W(x1 ,y1 ) and there would exist a neighborhood V(x1 ,y1 ) of p such that
V(x1 ,y1 ) W(x1 ,y1 ) S = . Then K0 K0 is covered by finitely many of the W(x1 ,y1 ) and the
intersection of the corresponding V (x1 , y1 )s is a neighborhood of p that does not intersect
K1 .
Because 2T z = T y + T y for all T , we get 2x0 = x1 + y1 , hence x0 = x1 = y1 , because x0
is an extremal point. But T x T y 6 V for all T G, hence x1 y1 6 V , and so x1 6= y1 .
This is a contradiction, which proves our initial assumption was false, and the conclusion
follows.
Proposition 2.5.1. The space X endowed with the weak topology is a locally convex
topological vector space.
The Hahn-Banach Theorem implies automatically that the weak -continuous linear func-
tionals on X separate the points of this space. In fact this can be seen directly. Each point
x X defines a weak -continuous linear functional x on X defined by
x () = (x).
Theorem 2.5.1. Every weak -continuous linear functional on X is of the form x for some
x X. Hence (X ) = X. 1
Proof. Assume that is a weak -continuous linear functional on X . Then | ()| < 1
for all in some set V (x1 , x2 , . . . , xn , ). This means that there is a constant C such that
| ()| C maxj |xj ()| for all X .
Let N be the set on which xj = 0, j = 1, 2, . . . , n. Then is zero on N , so we can
factor X by N so that we are in the finiteP dimensional situation. We can identify X /N
with Span(x1 , x2 , . . . , xn ) . In X /N , = j j xj , and so this must be the case in X as
well. Hence
X
= ( j xj ) ,
j
Kx = { C | || 1}.
1
It is important that on X we have the weak topology, if we put a different topology on it, (X ) might
not equal X.
2.5. THE DUAL OF A TOPOLOGICAL VECTOR SPACE 35
The set
Y
Kx
xV
We will show
(1) (K) is closed.
n (x + y) = n (x) + n (y).
For n large enough, n (x + y) approximates well (x), while n (x) and n (y) approximate
(x) and (y). By passing to the limit n we obtain that is linear.
Also for x V , |n (x)| 1, and again by passing to the limit, |(x)| 1. This implies
the continuity of , as well as the fact that it lies in (K). This proves (1).
For (2), note that is one-to-one, hence it is an inclusion. Moreover, the weak topology
was chosen so that it coincides with the topology induced by the product topology. Hence
(2).
so
So X has two topologies the one induced by the norm, and the weak topology. It is
not hard to check that the second is weaker than the first.
Here is an example.
Theorem 2.5.3. Let p [1, ). Then (Lp ([0, 1])) = Lq ([0, 1]), where q satisfies 1/p+1/q =
1. A function g Lq ([0, 1]) defines a functional by
Z 1
g (f ) = f (x)g(x)dx.
0
Proof. Note that every g Lq ([0, 1]) defines a continuous linear functional by the above
formula because of Holders inequality:
kf gk1 kf kp kgkq .
Moreover, it is not hard to see that if g1 and g2 define the same functional then g1 = g2
almost everywhere. This follows from the fact that if
Z 1
f (x)[g1 (x) g2 (x)]dx = 0
0
: A 7 (A )
is a measure on the Lebesgue measurable sets in [0, 1]. Note that is absolutely continuous
with respect to the Lebesgue measure, since if the Lebesgue measure of A is zero, then A
is the zero vector in Lp ([0, 1]), and so (A) = (A ) = 0.
Using the Radon-Nikodym Theorem, we deduce that there is a function g L1 ([0, 1])
such that
Z 1
(A ) = A (x)g(x)dx.
0
Approximating the functions in Lp ([0, 1]) by step functions, and using the continuity of both
the left-hand side on Lp ([0, 1]) and of the right-hand side on L ([0, 1]) Lp ([0, 1]), we
deduce that
Z 1
(f ) = f (x)g(x)dx for all f L ([0, 1]).
0
We also have
Z 1 1/p
1/p
kh kp = h(x)dx ,
0
Moreover kk = kgkq .
Theorem 2.5.5. (C([0, 1])) is the set of finite complex valued measures on [0, 1].
Proof. Each finite measure defines a continuous linear functional by
Z 1
(f ) = f (t)d.
0
Let us prove conversely, that every linear functional is of this form. For every complex
continuous linear functional , we have = Re + iIm where the real and the imaginary
part are themselves continuous. So we reduce the problem to real functionals. We show that
38 CHAPTER 2. LINEAR FUNCTIONALS
each such functional is the difference between two positive functionals, and then apply the
Riesz Representation Theorem.
For f 0, set
= + (+ ),
and the claim is proved. We can therefore write every continuous complex linear functional
as
= 1 2 + i(3 4 ),
= 1 2 + i(3 4 ).
Remark 2.5.1. Using the general form of the Riesz Representation Theorem, we see that
[0, 1] can be replaced by any compact space.
Theorem 2.5.6. (Banach-Alaoglu) Let X be a normed vector space. Then the closed unit
ball in X is weak -compact.
Proposition 2.5.3. Place a number in [0, 1] at each node of the lattice Z2 such that the
number at each node is the average of the four numbers at the closest nodes. Then all
numbers are equal.
2.5. THE DUAL OF A TOPOLOGICAL VECTOR SPACE 39
Proof. Consider the Banach space L (Z2 ). Let K be the set of elements in L (Z2 ) satisfying
the condition from the statement. Then K is a weak -closed subset of the unit ball; by
applying the Banach-Alaoglu theorem we deduce that it is weak -compact. It is also convex.
By the Krein-Milman Theorem, K = co(Ext(K)). Let f : Z2 [0, 1] be an extremal point in
K. Let L, U be the operators that shift up and left. Then U f, U 1 f, Lf, L1 f are functions
with the same property, and
1
f = (U f + U 1 f + Lf + L1 f ).
4
Because f is extremal, f = U f = U 1 f = Lf = L1 f , meaning that f is constant. In
fact f = 0 or f = 1. The convex hull of the two extremal constant functions is the set of
all constant functions with values in [0, 1], this set is closed, so K consists only of constant
functions. Done.
Theorem 2.5.7. The space L1 (R) is not the dual of any normed space.
Proof. If L1 (R) were the dual of a normed space, then the Banach-Alaoglu Theorem implies
that the closed unit ball in L1 (R) is weak -compact. By the Krein-Milman Theorem it has
extreme points. But this is not true, since every function in the closed unit ball can be
written as the convex combination of two functions in the unit ball.
Consider C([0, 1]), the Banach space of continuous functions on [0, 1], with the norm
kf k = supx[0,1] |f (x)|.
Theorem 2.5.8. (Stone-Weierstrass) Let A C([0, 1]) be a subalgebra with the following
properties
(1) if f A then f A,
Let us prove the claim. Suppose there is f A not constant on the support of . We
have f = f1 + if2 , so f1 = (f + f )/2 and f2 = (f f )/2i, and because f A, f1 , f2 A as
well. One of these is nonconstant, so we may assume that f is real valued. Replacing f by
(f + A)/B we may assume 0 f 1. Define the measures 1 and 2 by
d1 = f d, d2 = (1 f )d.
Then = 1 + 2 . Note
R 1 that 1 , 2 are
R 1both zero on A. R
We have k1 k = 0 f d||, k2 k = 0 (1 f )d||. And also kk = d|| = k1 k + k2 k.
Then
k1 k kk k2 k kk
= 1 + 2 .
kk k1 k kk k2 k
Note that
kk kk
1 , 2 K
k1 k k2 k
T : X Y.
Definition. A linear operator is called bounded if it maps bounded sets to bounded sets.
ker(T ) = {x X | T x = 0}.
Both ker(T ) and im(T ) are vector spaces. If T is a continuous linear operator between
topological vector spaces, then ker(T ) is closed. This is not necessarily true about im(T ).
41
42CHAPTER 3. FUNDAMENTAL RESULTS ABOUT BOUNDED LINEAR OPERATORS
These sets are convex and balanced. They are also closed, so by the Baire Category Theorem
is n such that the interior of Xn is nonempty. Because Xn is convex and balanced, its interior
contains the origin. Hence there is a ball B0,r centered at origin such that kT xk n for
all T F and x with kxk r. We have kT k n/r for all T F, and the theorem is
proved.
Here is an application that I learned from Hari Bercovici. We have
1 X
= (1)n1 xn1 .
x + 1 n=1
3.2. THE THREE FUNDAMENTAL THEOREMS 43
The left-hand side takes the value 1/2 when x = 1, so it is natural to impose that the
right-hand
Pn side converges to 1/2. A way to do this is to consider the sequence sn =
k1 k1
k=1 (1) x and then notice that
1
(s1 + s2 + + sn ) (3.2.1)
n
converges to the same limit as sn when the latter converges (Cesaro), but moreover for x = 1
(3.2.1) converges to 1/2.
Definition. A summation method associates to each convergent sequence sn , n 1 another
convergent sequence n , n 1 such that
(1) limn n = limn sn ;
(2) n =
P
k=1 nk sk for n = 1, 2, . . ., where nk is an array of complex numbers that does
not depend on sn and defines the summation method.
An example of a summation method, introduced by Cesaro, is nk = 1/n, n = 1, 2, . . .,
1 k n, and nk = 0 otherwise.
Theorem 3.2.4. (Toeplitz) The array nk , n, k 1, defines a summation method if and
only if it satisfies the following three conditions
(1) limn nk = 0, k = 1, 2, . . .;
(2) limn
P
k=1 nk = 1;
(3) supn
P
k=1 |nk | < .
Proof. Let us prove that the three conditions are necessary. If sn = nk for some k, then
n = nk . The fact that sn 0Pimplies limn nk = 0, hence (1).
If sn = 1, n 1, then n =
P
k=1 nk . Because sn 1, it follows that limn k nk =
1, hence (2).
For (3) we apply the Banach-Steinhaus Theorem. Denote by C0 the Banach space of
convergent sequences with the sup norm (i.e. continuous functions on N {} with the sup
norm, where N {} is given the topology such that the map f (x) = 1/xP from it to R is
a homeomorphism onto the image). Let n , n 1 be a sequence P such that n=1 converges
for every convergent sequence xn , n 1. We claim that n=1 |n | < . P
Indeed, if this is not the case, then choose rn > 0 such P that that Prn 0 and |n |rn =
. The sequence xn = rn n /|n | converges to 0, but n n xn = |n |rn = , which is
impossible. This proves our claim.
Additionally,
X X
sup n xn = |n |.
{xn }C0 ,kxn k1
n
n=1
The fact that the left-hand side does not exceed the right-hand side follows from the triangle
inequality. On the other hand, if xn = n /|n |, 1 n N and zero otherwise makes
44CHAPTER 3. FUNDAMENTAL RESULTS ABOUT BOUNDED LINEAR OPERATORS
n xn = N
P P
n=1 |n |. Taking N we obtain that the right-hand side is less than or
equal to the left-hand side. Hence the two are equal.
Define
X
n : C0 C, n ({sk }) = nk sk .
k=1
The sequence n ({sk }), n 1 is bounded for all convergent sequences {sk }. Hence by the
Banach-Steinhaus Theorem, kn k, n 1 is bounded, which is (3).
Now let us check that the conditions are sufficient. Let M = supn
P
k=1 |nk |. Consider
a sequence sn converging to s. We want to show that n converges to s as well. We compute
!
X X X
|n s| nk sk nk s + nk 1 s
k=1
k=1 k=1
X X
|nk ||sk s| + nk 1 |s|
k=1 k=1
XN X
|nk ||sk s| + M sup |sk s| + nk 1 |s|.
kN +1
k=1 k=1
We obtain limn |n s| = 0.
A = {T x | kxk < 1}
is a neighborhood of 0 in Y . We have
Y =
n=1 nA.
Because Y is of the second category (by the Baire Category Theorem), it follows that there
is n such that nA has nonempty interior. Consequently A has nonempty interior.
But A is convex and balanced, because it is the image through a linear map of a convex
and balanced set. Hence so it A, and consequently A contains a neighborhood of 0. Let
> 0 be such that
Fix y Y , kyk < and fix 0 < < 1. Choose x1 in the unit ball of X such that kyT x1 k < .
There is x2 X, kx2 k < / with ky T x1 T x2 k < 2 , ..., there is xn with kxn k < n1 /
and ky PT x1 T x2 T xn k < n . Because X is Banach, there is a point x X such
that x = n=1 xn . We have
ky T xk = lim ky T x1 T x2 T xn k = 0,
n
{(x, f (x)) | x A} A B.
1 : GT X, 1 (x, T x) = x
and
2 : GT Y, 2 (x, T x) = T x.
Both these operators are linear and continuous. The operator 1 is invertible and bijective.
By the Inverse Mapping Theorem (Theorem 3.2.6) its inverse is also continuous. We have
T = 2 11 , and hence T is invertible.
Here is an application.
46CHAPTER 3. FUNDAMENTAL RESULTS ABOUT BOUNDED LINEAR OPERATORS
Proof. Assume that the kernel and the image are closed. Because x = P x + (x P x), and
P (x P x) = P x P 2 x = 0 every element in X is the sum of an element in kerP and an
element in imP . Moverover, if x kerP imP , then x = P y, for some y, so P 2 y = P x = 0.
But P 2 y = P y = x, so x = 0. It follows that
X = kerP imP.
Let us show that the graph of P is closed. Consider a sequence (xn , P xn ), n 1, that
converges to (x, y); we want to show that y = P x. Because imP is closed, y imP . The
sequence xn P xn converges to xy. Because xn P xn kerP , there is z kerP such that
xn P xn z. So x y = z. It follows that x y kerP . We thus have P (x y) = P z = 0.
But But P y = y, so P (x y) = P x y. It follows that P x = y. From the Closed Graph
Theorem it follows that P is continuous.
Conversely, if P is continuous, then kerP = P 1 (0) is closed. Also, imP = ker(1 P ),
and 1 P is also continuous. Hence imP is closed.
Example. Let A be a closed subset of [0, 1], and let CA ([0, 1]) be the set of continuous
functions that are zero on A. Then there is a closed subspace Y of C([0, 1]) such that
Indeed, there is a bounded linear operator T : C(A) C([0, 1]) such that T g|A = g
(the complement of A is a disjoint union of open intervals, and on such an interval (a, b)
we can define T g(ta + (1 t)b) = tg(a) + (1 t)g(b)). If R : C([0, 1]) C(A) is the
restriction operator, then P = T R is a projection. It is also continuous because T and R
are continuous. Hence
Set Y = imP .
The operator T defined in this example is called a simultaneous extension. It has been
proved that such operators exist in more general situations (e.g. for compact spaces). The
existence of such an operator is a stronger version of the Tietze Extension Theorem.
Proof. We have
(1 + 2 ) T = 1 T + 2 T
kxk = sup{|(x)| | X , kk 1}
Proof. We have |(x)| kkkxk, so the left-hand side is greater than or equal to the right-
hand side. For the converse inequality, define 0 : Rx C, 0 (tx) = tkxk. Then k0 k = 1.
By the Hahn-Banach Theorem, there is a continuous linear functional : X C such that
kk = 1, and (x0 ) = kx0 k.
Proof. We have
and
Corollary 3.3.1. ker(T ) is weak closed, im(T ) is dense if and only if T is injective, and
T is injective if and only if im(T ) is weak dense.
Proof. Suppose (a) holds. Then by Proposition 3.3.1, (x) = 0 for all im(T ) if and only
if x ker(T ). We claim that the functionals that are zero on ker(T ) are the weak closure
of im(T ). Indeed, this set is weak closed and contains im(T ). To prove the converse
inclusion, recall that the dual of X with the weak topology is X. Assume that there is
0 that is zero on ker(T ) but is not in the weak -closure of im(T ). Then by the Hahn-
Banach Theorem, there is x X such that 0 (x) 6= 0 and (x) = 0 for all im(T ).
But (T x) = 0 for all Y means that T x = 0, so x ker(T ). Then 0 (x) = 0, a
contradiction. This proves our claim.
We are left to show that any functional that is zero on ker(T ) is in the image of T . Let
be such a functional. Define a linear functional on im(T ) by
(T x) = (x).
It is not hard to see that is well defined. Apply the Open Mapping Theorem to
T : X im(X)
to conclude that there is C > 0 such that for every y im(T ) there is x X such that
T X = y and kxk Ckyk. Hence
|(y)| = |(T x)| = |(x)| kkkxk Ckkkyk.
Hence is continuous. Extend to the entire space using Hahn-Banach. Because
(x) = (T x) = (T )(x),
it follows that = T . Hence im(T ), as desired. We thus proved that (a) implies (b).
(b)(c) is straightforward.
Now let us suppose that (c) holds. Let Z bet he closure of im(T ) in Y . Define S : X Z,
Sx = T x. As a corollary to Proposition 3.3.1, S : Z X is one-to-one.
If Z , the Hahn-Banach Theorem provides an extension Y of . For every
x X, we have
(T )(x) = (T x) = (Sx) = (S )(x).
Hence S = T . It follows that S and T have identical images, in particular the image
of S is closed. Apply the Inverse Mapping Theorem to S : Z im(S ) to conclude that
it is invertible. The conclusion follows from the following result.
Lemma 3.3.2. Suppose S : X Z is a bounded linear operator such that S : Z X
is invertible. Then S is onto.
Proof. Because S is invertible, there is C > 0 such that kk CkS k for all Z .
Let BX and BZ be the unit balls in X and Z. We will show that BZ CS(BX ), namely
that BZ S(BX ), where = 1/C.
Choose z0 6 S(BX ). Because S(BX ) is convex, closed, and balanced, an application of
the Hahn-Banach Theorem shows that we can separate it from z0 , so there is Z such
that k(z)k 1 for z S(BX ) but |(z0 )| > 1. If x BX , then
|S (x)| = |(Sx)| 1.
3.4. THE ADJOINT OF AN OPERATOR ON A HILBERT SPACE 49
Hence kS k 1. We have
< |(z0 )| kkkz0 k kz0 kkS k kz0 k.
We deduce that if kzk then necessarily z S(BX ).
Now let us show that moreover z S(BX ). Rescaling S we may assume = 1. Then
BZ T (BX ), and hence for every z Z and every > 0 there is x X such that kxk kyk
and ky T xk < . Choose z1 BZ . Let n = 31n (1 kz1 k). Define the sequences xn and
zn inductively as follows. Assume zn is already picked, and let xn be such that kxn k kzn k
and kzn PT xn k < n . Set zn+1
P = zn P T xn .
If x = xn , the T x = T xn = (yn yn+1 ) = z1 . Hence z1 T (BX ). This proves
our claim. The conclusion follows.
Using the lemma we conclude that im(S) = im(S), and so the image of S is closed. But
im(S) = im(T ), and so the theorem is proved.
Theorem 3.3.3. Let T : X Y be a bounded linear operator between Banach spaces.
Then im(T ) = Y if and only if T is one-to-one and im(T ) is norm closed.
Proof. By Proposition 3.3.1 T is one-to-one. By the Open Mapping Theorem there is
> 0 such that
{y Y | kyk } {T (x) | kxk 1}.
Then for a functional ,
kT k = sup{|(T )(x)| | kxk 1} = sup{|(T x)| | kxk 1}
sup{|(y)| | kyk } = kk.
We claim that given this inequality, im(T ) is closed. Let im(T ), and let n be such
that T n . Then T n is Cauchy, so the above inequality implies n is Cauchy. If is
its limit, then T = . The implication is proved.
By Theorem 3.3.2, im(T ) is closed, and by Proposition 3.3.1 it is dense. Hence
im(T ) = Y .
Lemma 3.4.1. Two linear operators S and T on a Hilbert space H are equal if and only if
(T ) = T.
(T + S) = T + S
(T ) = T
(ST ) = T S .
kT T k = kT k2 .
Proof. We have
kT T k kT kkT k = kT k2 .
normal if T T = T T
self-adjoint if T = T
unitary if T T = T T = I
an isometry if T T = I
hV x, V yi = hx, yi .
kT xk2 = hT x, T xi = hT T x, xi = hT T x, xi
= hT x, T xi = kT xk2
For the converse, note that all equalities in this sequence are obvious except hT T x, xi =
hT T x, xi. By Lemma 3.4.1 this condition is equivalent to the fact that T is normal.
Proof. The first property is a consequence of Proposition 3.4.3 and Proposition 3.4.2.
Assume that T is invertible. Then by the Inverse Mapping Theorem the inverse of T is
continuous, so we can choose = kT 1 k.
For the converse, the existence of such a implies that ker(T ) = {0}. Moreover, im(T )
is closed. And by the first property im(T ) is dense. So T is one-to-one and onto, hence
invertible.
Proposition 3.4.5. An operator A is self-adjoint if and only if hAx, xi is real for all x H.
52CHAPTER 3. FUNDAMENTAL RESULTS ABOUT BOUNDED LINEAR OPERATORS
Proof. If A is self-adjoint, then hAx, xi = hx, Axi. But by the properties of the inner product,
hx, Axi = hAx, xi. Hence the quantity must be real. Conversely, if the quantity is real then
So A = A by Lemma 3.4.1.
A projection P is called orthogonal if im(P ) = ker(P ) .
Proposition 3.4.6. A projection P is orthogonal if and only if P is self-adjoint.
Proof. Assume that P is orthogonal. Then every x H is of the form x = y + z with
y ker(P ) and z im(P ). Then
hP x, xi = hz, y + zi = kzk2
and
hx, P xi = hy + z, zi = kzk2 .
Hence P = P .
For the converse, note that P = P implies P normal, so ker(P ) = im(P ) = im(P ) .
But P is a projection, so im(P ) is closed. The conclusion follows.
As a corollary, a property that characterizes orthogonal projections is hP x, xi = kP xk2 .
Proposition 3.4.7. Let N be a normal operator. Then there are self-adjoint operators A1
and A2 that commute such that N = A1 + iA2 .
Proof. A1 = (N + N )/2, A2 = (N N )/2i.
Proposition 3.4.8. Let A be a self-adjoint operator. Then exp(iA) is unitary.
Proof. First, note that
iA A2 iA3
exp(iA) = I + +
1! 2! 3!
is well defined because the series is absolutely convergent hence convergent. We have
exp(iA) = exp(iA),
M T = T N.
Then
M T = T N .
exp(M )T = T exp(N ).
It follows that
T = exp(M )T exp(N ).
Set U1 = exp(M M ), U2 = exp(N N ). In view of the above corollary, these are unitary,
in particular kU k1 = kU2 k. We then obtain
Now replace M and N by M and N and repeat the same argument to conclude that
f () = exp(M )T exp(N ).
namely that fx,y is bounded. By Liouvilles theorem fx,y is constant. It follows that f itself
is constant, so f () = f (0) = T for all . Hence
exp(M )T exp( N ) = f () = T.
54CHAPTER 3. FUNDAMENTAL RESULTS ABOUT BOUNDED LINEAR OPERATORS
Write this as
exp(M )T = T exp(N ).
which must be equal term-by-term. Considering the -term we obtain that for all x, y,
hM T x, yi = hT N x, yi .
Hence M T = T N , as desired.
Corollary 3.4.2. If N is normal and T commutes with N , then T commutes with N and
N commutes with T .
Show that the hypothesis of the theorem does not necessarily imply M T = T N .
Chapter 4
Example. The Banach algebra B(X) of bounded linear operators on a Banach space X.
Example. The Banach algebra of continuous functions C([0, 1]).
We will almost always be concerned with Banach algebras over the complex numbers.
Definition. A series
X
c n an
n=0
55
56 CHAPTER 4. BANACH ALGEBRA TECHNIQUES IN OPERATOR THEORY
1 + b + b2 + b3 +
Hence
a1 = (1 b)1 = 1 + b + b2 + b3 + .
Definition. For a Banach algebra B, let G, Gr , and Gl be respectively the sets of invertible
elements, right invertible elements that are not invertible, and left invertible elements that
are not invertible.
Proposition 4.1.2. If B is a Banach algebra, then each of the sets G, Gr , and Gl is open.
then
a 7 a1
is continuous.
4.2. SPECTRAL THEORY FOR BANACH ALGEBRAS 57
Proof. Fix a G. We want to show that for every > 0, there is > 0 such that if
kb ak < then kb1 a1 k < . We have
So we have to choose so that kb1 k < /ka1 k for all b such that kb ak < . If
kb ak < 1/(2ka1 k), then k1 a1 bk < 1/2 and so by Theorem 4.1.1
1
kb1 k kb1 akka1 k = k(a1 b)1 kka1 k ka1 k = 2ka1 k.
1 21
B (a) = { C | a 6 G} .
B (a) = { C | a G} .
So the spectrum consists of those for which a is not invertible, and the resolvent
is the complement in C of the spectrum. When there is no risk of confusion, we ignore the
subscript, but be careful, the spectrum depends on the algebra in which your element lies
(in case the given element can be put inside several Banach algebras).
58 CHAPTER 4. BANACH ALGEBRA TECHNIQUES IN OPERATOR THEORY
Example. If B = Mn (C), the algebra of n n matrices, and A Mn (C), then (A) is the
set of eigenvalues.
Proof. First, note that if || > kak, then by Theorem 4.1.1 1 a/ is invertible. Hence
(1 a/) = a is invertible. This shows that the spectrum is included in the closed disk
of radius kak centered at the origin.
Let us show that the spectrum is nonempty. Assume to the contrary that for some
element a the spectrum is empty. Let be a continuous linear functional on B. Consider
the function
f : C C, f () = ((a )1 ).
(a 0 )1 [(a 0 ) (a )](a )1
f () f (0 )
lim = lim
0 0 0 0
1 1 2
= ( lim (a 0 ) (a ) ) = (a 0 ) .
0
(1 a/)1
<
1
.
1 ka/k
Hence
1
lim |f ()| = lim | (a/ 1)1
|| ||
1 1 1
limsup kkk(a/ 1)1 k limsup kk
|| || 1 ka/k
where for the last step we used Theorem 4.1.1. This last limit is zero. Hence f is a bounded
holomorphic function. By Liouvilles Theorem it is constant.
Using the Hahn-Banach Theorem we deduce that 7 (a )1 is constant, and since
the inverse is unique, it follows that 7 a is constant. But this is clearly not true.
Hence our assumption was false, and the spectrum is nonempty.
Since the map (a ) is continuous, and G (the set of invertible elements) is open,
the inverse image of G through this map is open. But the inverse image of G is the resolvent.
Hence the resolvent is open, and therefore the spectrum is closed. Being bounded (as it lies
inside the disk of radius kak), it is compact.
Proof. Fix an element a B and let || > kak. Then using Theorem 4.1.1 we can write
( a)1 = 1 + 2 a + 3 a2 + .
The series converges absolutely on every circle C(0, r) centered at the origin and radius
r > kak. We can therefore integrate term by term and write
Z
1
n
a = n ( a)1 d, n = 1, 2, 3, . . . (4.2.1)
2i C(0,r)
Let M (r) be the maximum of k(a)1 k on C(0, r), (which is finite because 7 (a)1
is continuous). Then
Hence
since again the series converges. The integral formula (4.3.1) would be meaningful only if in
this particular situation the two versions coincide. And indeed, we have the following result.
4.3. FUNCTIONAL CALCULUS WITH HOLOMORPHIC FUNCTIONS 61
we wish. Then we can ignore these sums and consider just the case where f (z) = N n
P
n=0 cn z .
To prove the result in this case, it suffices to check it for f a power of z. Thus let us show
that
Z
1
n
a = z n (z a)1 dz.
2i
Now we can rely on Cauchys theorem about the integral of a holomorphic function, to to
make a circle of radius greater than kak. Because on |z| > kak, we can expand
X
(z a)1 = ak /k k+1 .
k0
The series on the right is absolutely convergent, so we can integrate term-by-term to write
Z Z
1 n 1
X 1 nk1
z (z a) dz = z dz ak .
2i k=0
2i
All of the integrals are zero, except for the one where k = n, which is equal to 2i. Hence
the result is an , as desired.
A slight modification of the proof yields the following more general result.
Proposition 4.3.2. Suppose R(z) = P (z)/Q(z) is a rational function with poles outside of
the spectrum of a. Then R(a) is well defined and Q(a) is invertible, and
R(a) = P (a)Q(a)1 .
Theorem 4.3.1. (The Spectral Mapping Theorem for Polynomials) Let P (z) be a polyno-
mial and a an element in B. Then
(P (a)) = P ((a)).
Proof. Let (a). Then
P (a) P () = (a )Q(a).
Because a is not invertible, neither is P (a)P (). Hence P () (P (a)). Consequently
P ((a)) (P (a)).
Let (P (a)), and let 1 , 2 , . . . , n be the roots of P (z) = . Then
P (a) = (a 1 )(a 2 ) (a n ).
Because P (a) is not invertible, there is k such that a k is not invertible. Then
k (a), and = P (k ) P ((a)). This proves (P (a)) P ((a)). The double inclusion
yields the desired equality.
62 CHAPTER 4. BANACH ALGEBRA TECHNIQUES IN OPERATOR THEORY
Theorem 4.3.2. Let D be a domain in C that contains (a). Endow the space of holomor-
phic functions on D, Hol(D), with the topology of uniform convergence on compact subsets.
Then the map Hol(D) B, f 7 f (a) is a continuous algebra homomorphism.
Proof. The only difficult step is multiplicativity. But we have multiplicativity for polynomi-
als, and hence for rational functions. By Runges theorem, every function in Hol(D) is the
limit of rational functions. By passing to the limit in fn (a)gn (a) = (fn gn )(a), we conclude
that multiplicativity holds in general.
Theorem 4.3.3. (The Spectral Mapping Theorem for Holomorphic Functions) Let f be a
holomorphic function in a neighborhood of the spectrum of a. Then
(f (a)) = f ((a)).
Proof. Let (a). Then as before f (z)f () = (z )g(z) with g a holomorphic function
with the same domain as f . By the previous theorem
f (a) f () = (a )g(a),
which cannot happen. So f (z) is zero for some z (a), that is f ((a)).
Example. If R is such that im(R) is finite dimensional, then R is compact. Such an operator
is said to be of finite rank.
Theorem 4.4.1. The set K(X) of compact linear operators on X is a closed two-sided ideal
of B(X).
Proof. Let K1 and K2 be compact operators. Then K1 (B0,1 ) and K2 (B0,1 ) are compact.
Then
and the latter is compact because is the image through the continuous map (x, y) 7 x + y
of the compact set K1 (B0,1 K2 (B0,1 ) X X. This proves that K1 + K2 is compact.
4.4. COMPACT OPERATORS, FREDHOLM OPERATORS 63
Also for every C, if K is compact then K is compact, because the image of the set
K1 (B0,1 ) through the continuous map x 7 x is compact.
Finally, if T B(X) and K K(X) then T (K(B0,1 )) is the image of a compact set
through a continuous map, so it is compact. It follows that T K(B0,1 ) lies inside a compact
set, so its closure is compact. So T K is compact.
On the other hand, T (B0,1 ) is a subset of B0,n for some n, so KT (B0,1 ) is a closed subset
of the compact set K(B0,n ), hence is compact. This proves that KT is compact.
We thus showed that K(X) is an ideal. Let us prove that it is closed. Let Kn , n 1,
be a sequence of compact operators that is norm convergent to an operator T . We want
to prove that T is compact. For this we use the characterization of compactness in metric
spaces: Every sequence contains a convergent subsequence.
Let xk , k 1 be a sequence of points in the unit ball of X. Let us examine the sequence
T xk , n 1. For every > 0, there is n() such that for n n(), kKn xk T xk k kKn T k
. For n n(),
The sequence Kn xk has a convergent subsequence, and so we can find a subsequence Tn xkm
such that kT xkm T xkr k < 3 for all m, r. Do this for = 1, then choose the first term of a
sequence yk to be xk1 . Inductively let = 1/k, and choose from the previous sequence xkm
a subseqence such that kT xkm T xkr k < 3 and let yk be the first term of this subsequence.
The result is a Cauchy sequence T yk , which therefore converges. We conclude that T is
compact.
Proof. (a) If im(K) is closed, then it is a Banach space. The Open Mapping Theorem
implies that the image of the unit ball is a neighborhood of the origin. This neighborhood
is compact, and this only happens if im(K) is finite dimensional.
(b) The operator K|ker(K I) is a multiple of the identity operator. This operator is
also compact. By (a) this can only happen if we are in a finite dimensional situation.
(c) The operator K cannot be onto.
Theorem 4.4.3. Let B be a Banach algebra and let M be a two-sided closed ideal. Then
B/M is a Banach algebra with the norm
Proof. Let us show first that k k is a norm. Clearly if [a] = 0 then a M so k[a]k
ka ak = 0. Now assume that k[a]k = 0. Then There is a sequence mn M such that
limn ka + mn k = 0. Since M is closed, it follows that a M, so [a] = 0. Thus k[a]k = 0
if and only if [a] = 0.
If a B and C, then
k[a]k = k[a]k = inf{ka + mk | m M} = || inf{ka + m0 k | m0 M} = ||k[a]k.
Also
k[a] + [b]k = k[a + b]k = inf{ka + b + mk | m M} = inf{ka + m + b + m0 k | m, m0 M}
inf{ka + mk | m M} + inf{ka + mk | m M} = k[a]k + k[b]k.
Thus k k is a norm.
Next, let us show that the norm satisfies the requirements from the definition of a Banach
algebra. First,
k[1]k = inf{k1 mk | m M} = 1,
where the equality is attained for m = 0, and one cannot have k1 mk < 1 for in that case
m must be invertible and hence cannot be an element of an ideal.
Secondly, for a, b B, we have
k[a][b]k = k[ab]k = inf{kab mk | m M} inf{k(a m1 )(b m2 ) | m1 , m2 M}
inf{ka m1 k | m1 M} inf{kb m2 k | m2 M} = k[a]kk[b]k.
Finally, let us show that B/M is complete. Showing that every Cauchy sequence is
convergent is equivalent to showing that every absolutely convergent series is convergent. It
is clear that the fact that every Cauchy sequence is convergent implies that every absolutely
convergent series is convergent. For the converse, let xn , n 1 be a Cauchy sequence. By
k
choosing a subsequence, P we may assume that |yn ym | 1/2 whenever n, m k. Set
xk = yk+1 P yk . Then xk is absolutelyPconvergent, and its sum is the limit of yk .
So let n [an ] be a series such that n k[anP ]k = M < . Then for each n there is mn
n
sucht that kan + mn k kan k + 1/2 . Hence n (an + mn ) is absolutely convergent, and
therefore convergent in B. If a is its sum, then a + M is the sum of the original series in
B/M. This concludes the proof that B/M is a Banach algebra.
Corollary 4.4.1. The algebra B(X)/K(X) is a Banach algebra.
Definition. The algebra B(X)/K(X) is called the Calkin algebra.
Definition. An operator with finite dimensional kernel and with closed image of finite
codimension is called Fredholm.
Theorem 4.4.4. (Atkinson) Let H be a Hilbert space. Then the Fredholm operators form
the preimage through the quotient map of the invertible elements of B(H)/K(H).
Corollary 4.4.2. The Fredholm operators form an open set.
Definition. If T is Fredholm, then the index of T is
ind(T ) = dim ker(T ) codim im(T ) = codim im(T ) dim ker(T ).
One can show that the index is continuous and invariant under compact perturbations.
4.5. THE GELFAND TRANSFORM 65
Proof. Since (a (a)) = 0 it follows that every element in B is of the form + a, for
some C and a ker(). Note that if 6= 0 and k + ak < || = |( + a)|, then a is
invertible. This cannot happen, because (a) = 0 implies 1 = (aa1 ) = (a)(a1 ) = 0.
Hence |(b)| kbk for all b. Because (1) = 1, the equality is attained, so kk = 1.
not in the closure of M. Thus the closure of M is an ideal, and because of maximality, this
ideal must be M. So M is closed.
The quotient algebra B/M is a division algebra, because M is maximal. So by the
Gelfand-Mazur Theorem it is C. The quotient map is the desired multiplicative functional.
Recall that for every a B, the function a : (B )1 C given by a() = (a) is continuous,
where (B )1 is the unit ball in B , endowed with the weak topology.
Definition. The Gelfand transform of the Banach algebra B is the function : B C(MB )
given by (a) = a|MB .
Theorem 4.5.1. The Gelfand transform is an algebra homomorphism and k(a)k kak
for all a B.
Proof. is clearly linear and (1) = 1. Let us check that is multiplicative. We have
If B is not commutative, the Gelfand transform has large kernel which is generated by
the elements of the form ab ba. For this reason it is not so interesting.
Proposition 4.5.4. If B is a commutative Banach algebra and a B, then a is invertible
in B if and only if (a) is invertible in C(MB ).
Proof. If a is invertible, then (a1 ) = ((a))1 . If a is not invertible, then M0 = {ab | b B}
is a proper ideal. It is contained in a maximal ideal, whose associated functional is zero on
a. Hence (a) is not invertible.
Remark 4.5.1. The fact that a invertible implies (a) invertible does not use the fact that
the Banach algebra is commutative. Because (abba) = (a)(b)(b)(a) = 0, it follows
that ab ba is not invertible. This means that the canonical commutation relations for the
position and momentum operators in quantum mechanics
~
P Q QP = I
i
cannot be modeled with bounded linear operators.
Proposition 4.5.5. If B is a commutative Banach algebra and a B, then B (a) = im((a))
and rB (a) = k(a)k.
Proof. If is not in (a), then a is invertible. This is equivalent to (a) is invertible.
And this is further equivalent to the fact that is not in the image of (a).
Chapter 5
C algebras
Definition. A C -algebra is a Banach algebra over the complex numbers with an involution
that satisfies
(a + b) = a + b
(a) = a
(ab) = b a
(a ) = a.
ka ak = kakka k. (5.1.1)
ka ak = kak2 . (5.1.2)
The two conditions (5.1.1) and (5.1.2) are equivalent, though it is hard to show that
(5.1.1) implies (5.1.2). Thus our working definition will be the one with (5.1.2), what is
usually called a B -algebra. This condition implies (5.1.1) as follows:
kxk2 = kx xk kxkkx k.
Example. The algebra B(H) of bounded linear operators on a Hilbert space with the invo-
lution defined by taking the adjoint.
67
68 CHAPTER 5. C ALGEBRAS
Example. Let X be a compact Hausdorff space. The algebra C(X) of complex valued
continuous linear functions on X with the sup norm and the involution given by (f )(x) =
f (x) is a C -algebra.
Example. The algebra K(H) of compact operators on a Hilbert space H is a C -algebra.
We know that it is a subalgebra of B(H), so all we have to check is that it is closed under
taking the adjoint. Thus we have to show that the adjoint of a compact operator is compact.
Let T be compact and consider a sequence yn , n 1 in the unit ball B0,1 centered at
the origin of H. Let us prove that T yn has a convergent subsequence. Define the functions
fn : T (B0,1 ) C,
fn (x) = hx, yn i
where h, i is the inner product. Note that since T is compact, the domain of these functions
is compact. Then
|fn (x)| kxkkyn k 1.
So kfn k, n 1 is a bounded sequence. Also,
|fn (x) fn (x0 )| kyn kkx x0 k kx x0 k,
so the sequence fn is also equicontinuous. By the Arzela-Ascoli theorem, it has a convergent
subsequence in C(T (B0,1 )). Note also that
Theorem 5.2.2. (The Spectral Theorem) If H is a Hilbert space and N is a normal operator
on H, then the C -algebra CN generated by N and N is commutative. Moreover, the
maximal ideal space of CN is homeomorphic to (N ) and hence the Gelfand map is a -
isometrical isomorphism of CN onto C((N )).
Proof. The algebra CN is commutative because it is the closure of the algebra of all polyno-
mials in N and N .
Let us show that the set of multiplicative functionals, MCN , is homeomorphic to (N ).
In view of Proposition 4.5.5, we can define the onto function
: MCN (N ), () = (N )().
(N ) = (N )() = (N )(0 ) = 0 (N ),
and also
B0 ,r = { (N ) | | 0 | < r}.
Set 0 = 1 (0 ). Then
: B B(H),
for some Hilbert space H, that is non-degenerate in the sense that (a)x is dense when a
ranges through a and x ranges through H. A vector x is called cyclic if the set {(a)x | a B}
is dense in H; in this case the representation is called cyclic.
Definition. A state on a C -algebra is a linear functional such that (a a) 0 for all a
and kk = 1.
Note that a state has the property that (1) = 1.
Theorem 5.3.1. (The Gelfand-Naimark-Segal Construction) Given a state of B, there is
a -representation : B B(H) which is cyclic, and a cyclic vector x such that
(a)b = ab.
ha, bi = (b a).
This has all the nice properties of an inner product, except that it might be degenerate, in
the sense that there might be a such that ha, ai = (a a) = 0. Adapting the Cauchy-Schwarz
inequality, we deduce that the set N of elements a such that ha, ai = 0 form a subspace of
B.
Let us show that N is also a left ideal of B. This is because of the Cauchy-Schwarz
inequality:
Then B/N is an inner product space. Consider the completion H of this space, which
is therefore a Hilbert space. We have
where
c = c = (kak2 a a)1/2 = (ka ak a a)1/2 .
The element c can by defined because the function f (t) = (ka ak a a)1/2 is continuous on
(a a), so we can use Theorem 5.2.2. So, because is positive,
It follows that
so (a) is continuous. This implies that (a) can be extended to the entire Hilbert space
H . This representation is cyclic, with cyclic vector 1 + N . Also, h (a)1, 1i = (1 a1) =
(a).
The set of states is a weak closed convex subset of the unit ball of B . The extremal
points are called pure states.
: B B(H ), =
where the sum is taken over all pure states. It suffices to show that is faithful, namely
one-to-one, because the fact that it is an isometric -homomorphsm then follows from The-
orem 5.4.3 proved in next section. For this we use the theorem of M. Riesz about extension
of positive functionals. Let a be a nonzero element of B. Then there is a state such that
(a a) > 0.
Consider the GNS representation associated to , and let x be its cyclic vector. Then
By the Krein-Milman theorem, there is a pure state that satisfies (a a) > 0. In this case
6= 0, hence the representation is faithful. The theorem is proved.
Definition. Let H be a Hilbert space. The weak operator topology is the topology defined
by the open sets
The strong operator topology is the topology defined by the open sets
Remark 5.4.1. If C is a self-adjoint subalgebra of B(H), then its weak closure is a von
Neumann algebra. If C is commutative, then its closure is commutative.
We want to show that there is a unique L space structure on MWN and a unique -
isometrical isomorphism : WN L ((N )) which extends the functional calculus with
continuous functions defined by the Spectral Theorem (Theorem 5.2.2).
Assume we have a finite positive regular Borel measure on MWN . We can assume that
the measure of the entire space is 1, so that we have a probability measure. For the moment,
we work in this hypothesis.
The map 7 Mf , where Mf : L2 (MWN ) L2 (MWN ), Mf g = f g identifies L (MWN )
with a maximal commutative von Neumann subalgebra of the algebra of operators on
L2 (MWN ).
Proposition 5.4.2. The weak operator topology and the weak topology on L coincide.
Proof. L = L1 , and recall that every function in L1 is the product of two L2 functions.
Thus an element of the form
Z
(f ) = f g
Proof. We will show that the unit ball in C((N )) is weak -dense in the unit ball in
P
L ((N )). Consider a step function in the unit ball of L , f = j Ej , |j | 1 with Ej
disjoint and their union is (N )). For each j, choose Kj Ej . Using Tietzes Extension
5.4. FUNCTIONAL CALCULUS FOR NORMAL OPERATORS 73
Theorem we can find g in the unit ball of C((N )) such that g(x) = j for x Kj . Then
for h L1 ,
Z Z
h(f g) |h||f g|
n Z
X n Z
X
= |h||f g| |h|
j=1 Ej \Kj j=1 Ej \Kj
Because the measure is regular, we can choose Kj such that the integrals are as small as
desired.
Recall that a vector x is cyclic for an algebra B B(H) if Bx is dense in H and separating
if T x = 0 implies T = 0. If B is commutative, then x cyclic implies x separating, because
T x = 0 implies Bx ker(T ), hence T = 0.
Theorem 5.4.1. If N is a normal operator on H such that CN has a cyclic vector, then
there is a positive regular Borel measure supported on (N ) = MCN and an isometrical
isomorphism from H onto L2 ((N ), ) such that the map
: WN B(L2 ((N ), ), (T ) = T 1
Proof. Let x be a cyclic vector for CN with kxk = 1. Consider the linear functional on
C((N )) defined by (f ) = hf (N )x, xi. Then is positive because if f 0 then f = g 2 for
some real valued function g, and then
and hence
We also have
If the support of were not the entire spectrum, then, by Urysohns lemma, we could find
a continuous function f that is 1 somewhere on the spectrum and is zero on the support of
74 CHAPTER 5. C ALGEBRAS
impossible. So supp() = (N ).
Define
0 : CN x L2 ((N ), ), 0 (f (T )x) = f.
The computation
Z
kf k22 |f |2 d = |f |2 (N )x, x = kf (N )xk2
=
(N )
shows that 0 is a Hilbert space isometry. Because CN is dense in H and C((N )) is dense
in L2 ((N ), ), 0 can be extended uniquely to an isometrical isomorphism
: H L2 ((N ), ).
Moreover, if we define
(f (N )) = Mf = (f (N )).
Because the weak operator topology and the weak topology coincide on L (Proposi-
tion 5.4.2), is a continuous map from WT with the weak operator topology to L ((N ), )
with the weak topology. And because continuous functions are weak -dense in L , it fol-
lows that (WT ) = L ((N ), ). Thus is a -isometrical isomorphism mapping WT onto
L ((N ), ).
Finally, if (1 , 1 ) are a different pair with the above properties, then 1 1 is a
-isometrical isomorphism from L ((N ), 1 ) onto L ((N ), ) which is the identity on
C((N )). Then and 1 are mutually absolutely continous, L ((N )) = L ((N ), 1 )
and 1
1 is the identity. This completes the proof.
5.4. FUNCTIONAL CALCULUS FOR NORMAL OPERATORS 75
However, not all operators have cyclic vectors. Instead we will use non-separating vectors
and replace H by the smallest invariant subspace containing a non-separating vector. We
proceed to show that every normal has a separating vector.
An easy application of Zorns lemma yields the following result.
Lemma 5.4.1. Let T B(H), V is a closed subspace of H, and PV the orthogonal projection
onto V . Then PV T = T PV if and only if V is an invariant subspace for both T and T .
Moreover, in this case both V and V are invariant for T .
Lemma 5.4.2. If B is a C -algebra contained in B(H) and v H, then the projection onto
Bv is in B 0 .
Proof. By Lemma 5.4.1, it suffices to show that Bv is invariant for both T and T for every
T B. Note that T B so both T and T leave Bv invariant.
E E0 = E0 E = 0 for 6= 0 .
76 CHAPTER 5. C ALGEBRAS
Clearly E is not empty, since we can build an element in E starting with one vector, via
Lemma 5.4.2. Order E by inclusion. The hypothesis of Zorns Lemma is satisfied. Pick a
maximal element {E }A .
Let F be the collection of all finite subsets of the index set A partially ordered by inclusion
and let {PF }F F be the net of the orthgonal projections defined by
X
PF = E .
F
0
Then the net is increasing. If F > F then
k(PF PF 0 )xk2 = (PF PF 0 )2 x, x = h(PF PF 0 )x, xi = hPF x, xi hPF 0 x, xi .
The net hPF x, xi is increasing and bounded from above by kxk2 , so it is convergent. Hence
it is Cauchy, and so is kPF xk. Then PF x is norm convergent. Define P x = limF PF x. Then
P is an orthgonal projection.
The range V of P has the property that both V and V are invariant under B. Moreover
P B by Lemma 5.4.2. Note that if v V , then Bv is orthogonal to each E so we can
add the projection onto this space to the family {E } , contradicting maximality. Hence
V = 0. P
Because H is separable, A is countable. Thus we can define w = v . Then for each
, the range of E is contained in Bw. So w is a cyclic vector for B.
Corollary 5.4.1. Every commutative C -algebra of operators on a separable Hilbert space
has a separating vector.
Theorem 5.4.3. If : B1 B2 is a -homomorphism of C -algebras, then kk 1 and
is an isometry if and only if it is one-to-one.
Proof. If a B1 and a = a (i.e. a is self-adjoint), then Ca is a commutative C -algebra con-
tained in B1 and (Ca ) is a commutative C -algebra contained in B2 . If is a multiplicative
linear functional on (Ca ), then is a multiplicative linear functional on Ca . Because of
the Gelfand-Naimark Theorem, we can choose so that |((a))| = k(a)k. Then
kak |((a))| = k(a)k,
so is a contraction on self-adjoint elements. For arbitrary b B1 ,
kbk2 = kb bk k(b b)k = k(b) (b)k = k(b)k2 .
Hence kk 1.
For the second part, clearly if is an isometry than it is one-to-one. Assume that is
not an isometry and choose b such that kbk = 1 but k(b)k < 1. Set a = b b; then kak = 1
but k(a)k = 1 with > 0. Choose a function f C([0, 1]) such that f (1) = 1 and
f (x) = 0 if 0 x 1 . Using the functional calculus on Ca , define f (a). Since
(f (a)) = im((f (a)) = f ((a)),
we conclude that 1 (f (a)), so f (a) 6= 0. We have (f (a)) = f ((a)) (true on polynomials
then pass to the limit). But k(a)k = 1 , so ((f (a)) [0, 1 ]. Hence (f (a)) =
f ((a)) = 0. Thus is not one-to-one.
5.4. FUNCTIONAL CALCULUS FOR NORMAL OPERATORS 77
Let H be a separable Hilbert space and N normal on H. By Corollary 5.4.1, the com-
mutative von Neumann algebra WN has a separating vector x. If we set Hx = WN x, then
both Hx and Hx are invariant under WN . We can therefore define a map : WN B(Hx )
by (N ) = N |Hx .
Lemma 5.4.3. The map defined above is a -isometrical isomorphism. Moreover B(H) (T ) =
B(Hx ) (T |Hx ) for all T WN .
Proof. In view of the previous theorem, let us show that is one-to-one. And indeed, if
(T ) = 0 then T x = 0, because x = Ix WN x. So T = 0, because x is separating of WN .
The equality of spectra is proved as follows.
First,
B(H) (T ) = WN (T ).
WN (T ) = (WN ) (T |Hx )
because is a -isometrical isomorphism onto the image. Repeating the above argument we
also have
Theorem 5.4.4. (Functional Calculus for Normal Operators - Version I) Let N be a normal
operator on the separable Hilbert space H let the : CN C((N )) be the Gelfand
transform. Then there is a positive regular Borel measure having support (N ) and a
-isometrical isomorphism from WN onto L ((N ), ) which extends . Moreover is
unique up to mutual absolute continuity while L ((N ), ) and are unique.
x : WN B(Hx ), x (T ) = T |Hx .
Let Wx be the von Neumann algebra generated by T |Hx . The map is continuous in the
weak operator topology (because it is obtained by restricting the domain). Hence (WN )
Wx . Moreover, if
0 : CN |Hx C((N ))
Because N |Hx is normal and has the cyclic vector x, by Theorem 5.4.1 there is a positive
regular Borel measure with support (N |Hx ) = (N ) (here we use the previous lemma),
and a -isometrical (onto) isomorphism
0 : Wx L (, (N ), ), such that 0 |CN |Hx = .
Moreover, 0 is continuous form the weak operator topology of Wx to the weak -topology on
L ((N ), ). Hence = 0 is a -isometrical isomorphism from WN into L ((N ), ),
continuous in the weak/weak topologies, and which extends the Gelfand transform.
The only thing that remains to show is that takes WN onto L ((N ), ). For this we
need the following result.
Lemma 5.4.4. Let H be a Hilbert space. Then the unit ball of B(H) is compact in the
weak operator topology.
Proof. The proof is from the book of Kadison and Ringrose, Fundamental of the theory of
operator algebras. For two vectors x, y H, let Dx,y be the closed disk of radius kxk kyk
in the complex plane. The mapping which assigns to each T (B(H))1 the point
Y
{hT x, yi | x, y H} Dx,y
x,y
is a homeomorphism of (B(H))1 with the weak operator Q topology onto its image X in the
topology induced on X by the product topology of x,y Dx,y . As the latter is a compact
Hausdorff topology by Tychonoffs theorem, X is compact if it is closed. So let us prove
that X is closed.
Let b X. Choose x1 , y2 , x2 , y2 H. Then for every > 0 there is T (B(H))1 such
that each of
|a b(xj , yk ) a hT xj , yk i , |b(xj , yk ) hT xj , yk i |,
|b(ax1 + x2 , yj ) hT (ax1 + x2 ), yj i |, |b(xj , ay1 + y2 ) hT xj , ay1 + y2 i |
is less than . It follows that
|b(ax1 + x2 , y1 ) a b(x1 , y1 ) b(x2 , y1 )| < 3
|b(x1 , ay1 + y2 ) a b(x1 , y1 ) b(x1 , y2 )| < 3.
Thus
b(ax1 + x2 , y1 ) = a b(x1 , y1 ) + b(x2 , y1 ) b(x1 , ay1 + y2 ) = ab(x1 , y1 ) + b(x1 , y2 ).
Additionally, |b(x, y)| kxk kyk. Hence b is a conjugate-bilinear functional on H bounded
by 1. Using the Riesz representation theorem we conclude that there is an operator T0 such
that b(x, y) = hT0 x, yi. This operator has norm at most 1 and we are done.
Using the lemma, we obtain that the unit ball in WN is compact in the weak operator
topology. It follows that its image is weak -compact in L ((N ), ), and hence weak -closed.
Since this image contains the unit ball of C((N )), it follows that it contains the unit ball
in L ((N ), ). Hence takes the unit ball in WN onto the unit ball of L ((N ), ). So
is onto.
The uniqueness is as in Theorem 5.4.1. We are done.
5.4. FUNCTIONAL CALCULUS FOR NORMAL OPERATORS 79
It turns out that f (N ) is the functional calculus defined by the Gelfand transform. In fact
muc more is true.
Let f : (N ) C be a measurable function. There is a countable collection of open
disks, Di , i 1, that form a basis for the topology on (N ). Let V be the union of those
disks Di for which (f 1 (Di )) = 0. Then (f 1 (D)) = 0. The complement of V is the
essential range of f . We say that f is essentially bounded if its essential range is bounded.
Theorem 5.4.5. (Functional Calculus for Normal Operators - Version II) There is a -
isometrical isomorphism : L ((N ), ) WN which is onto, defined by the formula
Z
h(f )x, yi = f dx,y .
(N )
Moreover, = 1 .
Proof. Check on step functions, then use density.
This justifies the notation
Z
f (N ) = f dE.
(N )
In particular,
Z
N= tdE.
(N )
80 CHAPTER 5. C ALGEBRAS
Theorem 5.4.6. (The spectral mapping theorem) The spectrum of f (N ) is the essential
range of f .
Proposition 5.4.6. If N is normal and has spectral measure EN , and if f L (EN ), then
f (N ) is also normal and the spectral measure of f (N ) is defined by Ef (N ) () = E(f 1 ()).
Multiplication operators
Distributions
81
82 CHAPTER 6. TOPICS PRESENTED BY THE STUDENTS
Appendix A
Background results
83